This paper advocates this new approach to dialogue management, based on a Constructive Dialogue Model. Dialogue is understood as a joint activity between ...
Constructive Dialogue Management: cooperation based on communicative principles
Kristiina Jokinen Computational Linguistics Laboratory Nara Institute of Science and Technology3 1 Introduction
When considering di erent dialogue systems, three approaches to dialogue management can be distinguished. Special purpose NL interfaces aim at providing an appropriate service with simple dialogues constrained by the requirements of a particular task. Usually the system reacts mechanically to the user's request, and the user is required to nd an appropriate question strategy to get the information she wants. Many information retrieval systems (library catalogues, ight information, tourist information) are of this type. The second approach aims at building conversing dialogue systems which allow active dialogue partners to engage in mixed initiative dialogues. This presupposes the modelling of realistic conversational settings, and thus requires explicit representation of the speakers' beliefs, goals and plans, as well as characterisation of appropriate dialogue behaviour by conversational roles. HAM-ANS [16] and KAMP [2] were early systems in this line, PEA [14] and EDGE [4] are recent examples. Recent advances in resource-bounded agency [5, 3], con ict resolution [8] and rational agency [7] have led to a third approach, focussing on conversational agents. The system models a communicating agent, albeit a simple one, which can reason about communicatively adequate behaviour. Rather than being hardcoded in the system's control structure, rationality and cooperativeness are regarded as general communicative principles which govern the whole reasoning process, and the system can thus adapt itself to di erent communicative situations. This paper advocates this new approach to dialogue management, based on a Constructive Dialogue Model. Dialogue is understood as a joint activity between rational and cooperative agents. It is initiated to accomplish some independently set real world task which cannot be achieved by the agents acting alone: to obtain required information, the agents need to communicate with each other. Since the agents act in interactive social situations, their behaviour is constrained by obligations which describe rational and cooperative action. On the one hand, there is a need or an obligation to react (quickly) to the changing environment 3
(open a window when it is too hot, nd food when hungry, reply when addressed, etc.). On the other hand, smoothness of interaction requires that agents should not prevent other agents from ful lling their goals, and thus actions which show the agents' sincerity, motivation and consideration towards the other agent(s) are favoured. Hence, by means of communication, the agent constructs a model of how to achieve her goals and simultaneously take into account contextual requirements and the partners' goals. Communicative contributions are reactions to the partner's previous contribution, evaluated in the changed context with respect to the communicative obligations. The paper is organised as follows. Section 2 describes the Constructive Dialogue Model (CDM) and its main features. Section 3 presents the system architecture. Section 4 deals with constructive planning and goal formulation. Finally, section 5 draws conclusions.
2
Constructive Dialogue Model
The Constructive Dialogue the following claims:
Model (CDM) is based on
1. Dialogue is cooperative negotiation rather than a simple question-answer sequence, cf. [14, 6]. 2. Participants are rational, cooperative agents [13]. 3. The agent with a need initiates the dialogue [10]. 4. The agents push their own goal forward while showing consideration to the partner's goal. 5. The agents exchange new information. 6. Contributions are reactions to the immediately previous contribution in the changed context. 7. Contributions are locally planned and realised: No
prede ned dialogue structure: coherence depends on the domain organisation. No prede ned dialogue acts: communicative intentions are encoded in expressive and evocative attitudes associated with each contribution. 8. Dialogue nishes when the agents agree that their goals have been successfully achieved.
I would like to thank Yuji Matsumoto for providing an A communication cycle consists of the analysis of the excellent environment to do research and Graham Wilcock for helpful discussions on the topic. partner's contribution, its evaluation in the changed
context with regard to the agent's own knowledge and intentions, and nally, reporting of the results of the evaluation (cf. [1]). The cycle produces new information which changes the context and obliges the partner to evaluate the change with respect to her knowledge, then report back the result. Dialogue is thus managed locally and collaboratively, and it nishes when both participants report that their goals are ful lled in the context. Communicative principles describe rational and cooperative action, and provide a uniform basis for planning both what to say and how to say it. They are encoded as inference rules which operate on the speakers' beliefs and intentions, and on contextual facts such as the central concept of the dialogue, new information to be communicated, initiatives and dialogue expectations, see [12].
partner's goal with respect to the context, and produce a reaction by specifying one's goal into a surface contribution. The Task Manager, however, tries to prove a given task goal (TGoal) in the current context. If the proof succeeds, the next goal in the goal stack is proved, but if it fails, a recovery strategy is invoked. Recovery strategies deal with task re-planning or asking help from the partner via DM. The Task Manager and the Dialogue Manager communicate with each other by sending requests and providing answers. TM initiates DM when it thinks that the partner may possess knowledge that is needed to accomplish the given task (claim 3). DM then produces a communicatively appropriate dialogue contribution and awaits the partner's response. The response is analysed and evaluated, and the next CGoal determined. The CGoal can be either to continue the dialogue on the basis of the current dialogue situation, or 3 The system architecture to tell TM to continue its reasoning. Figure 2 shows The CDM system1 consists of a Dialogue Manager the communication between DM and TM. (DM), a Task Manager (TM), and their knowledge bases, as shown in Figure 1. The key resource is the Context Model, a dynamic knowledge base containing information about the mental state of the participants (beliefs, expectations, wants, intentions etc.) as well as the current dialogue and task situation (contributions, topic, new information, dialogue goals, task goals etc.). DM also has access to Communicative Principles (rules and preferences that describe rational and cooperative communication) and World Model (general knowledge about entities and their relations in the world), while TM has access to Application Model (structure of the domain). The Application Model is assumed to be subsumed by the World Model, and thus DM can use wider knowledge than TM in its reasoning about approriate continuations. Figure 2: Dialogue Manager and Task Manager. PARTNER
Task Manager(TGoal)
Dialogue Manager(DGoal)
prove(TGoal)
interpret(DGoal)
nextGoal(TGoal,NextG)
evaluate(DGoal,CGoal)
recover(TGoal)
askTM(CGoal,TGoal) tellTM(CGoal,TGoal)
assume(TGoal,NextG) askDM(TGoal,CGoal)
TaskManager(NextG)
react(CGoal)
The di erence in the operation of DM and TM is based on the fact that dialogue actions are di erent Dialogue Task Application Manager Manager Model from actions in general. As Grosz and Sidner [9] argue, (DM) (TM) dialogue actions need to be recognised, i.e. the speaker World Model intends the hearer to recognise her communicative plan (intentions behind the dialogue act), which is not the case in general theories of action. Moreover, general actions can be divided into subactions, the performance Context Model of which constitutes an act, while this is not the case with dialogue acts. In dialogues, several contributions may constitute an act or a contribution may contain Figure 1: The system architecture. several acts, which requires constructive evaluation of The Dialogue Manager operates in three phases, cor- the goal rather than step-wise planning. responding to the three reasoning processes that agents undertake in communication: accept and interpret the 4 Constructive planning partner's goal (DGoal), choose the dialogue strategy A cooperative dialogue partner tries to react in the way and formulate one's own goal (CGoal) by evaluating the her partner intended to evoke. However, if the system always adopts the user's evocative intentions, it can 1 The prototype is implemented in SICStus Prolog. only provide simple answers to clear and unambiguCommunicative Principles
response expected
central concept related unrelated
non-expected
related unrelated
goals unful lled ful lled unful lled ful lled unful lled ful lled unful lled ful lled
speaker initiative backto nish/start repeat-new, X nish/specify subquestion, X continue/start object, X object/specify
partner initiative follow-up-old follow-up-new new question new request continue somethingelse notrelated new indir-request
Figure 3: Possible dialogue strategies for binary valued contextual features. X marks con ict situations, where the speaker must decide whether to persist on own goal or take up the partner's goal. ous user requests. As claimed above (claim 1), this is a simpli ed view of dialogues, which are joint activities between rational agents. Moreover, [8] points out that the system is benevolent rather than cooperative, if it always adopts the partner's goal simply because the partner has the goal, and thus abstracts away from con ict situations by constraining the context in a particular way. Given that the partner's goal is recognised, the rst step in the evaluation of the goal is to nd an appropriate dialogue strategy. We call the strategy joint purpose as it is chosen on the basis of contextual knowledge so that it both advances the agent's own goal and takes the partner's goal into consideration (claim 4).2 On the basis of our corpus studies3 , this reasoning is found to depend on contextual features such as whether the contribution ful ls expectations, whether it is thematically related to the previous dialogue, whether the speaker has the initiative and whether the speaker has unful lled goals. Two of the rules that encode the system's rationality in determining a joint purpose can be verbally represented as follows (the italicized responses indicate the contributions planned when the constraints hold): (1) If the partner's contribution is expected and thematically related, and if the speaker has the initiative, then the speaker can go back to her previous goal and push it forward, given that the goal is not ful lled in which case the dialogue can be closed.
A1: B1: A2:
B2: 2
I want to rent a car. Where? In Bolton.
Ok. Here is a list of car hire companies in Bolton:
Dialogue participants may have goals which are not shared by their partners. However, they must have at least one common goal for the communication to take place at all, and the joint purpose encodes this. 3 The corpus [15] consists of simulated human-computer dialogues where the user needs to nd information from car-hire companies and restaurants in a particular area.
(2) If the partner's contribution is not expected nor the-
matically related, and the speaker has the initiative and unful lled goals, then the speaker can repeat the Central Concept of the previous unful lled goal.
A1: B1: A2:
B2:
I need a car. Do you want to buy or rent one? Do you know any restaurants?
Oh, I thought we were talking about cars, and wanted to know if you want to buy or rent a car.
The current prototype system has 16 joint purposes, based on binary valued contextual features (see [11]). Figure 3 summarises the joint purposes and their contextual conditions. After choosing the strategy, the agent formulates a communicative goal (CGoal) which she wants to achieve. Although communication itself arises from the need or wish to make the partner provide necessary information for the accomplishment of a given task, each CGoal is formulated on the basis of communicative knowledge: new information is exchanged so that communicative requirements are ful lled (claims 5,6). The repertoire of communicative strategies depends on the agent's knowledge and experience ( exibility, creativity), but a greater set does not automatically guarantee a greater communicative capability. It provides a greater initial variety for responses but the agent can compensate her limited strategies by reasoning with respect to communicative obligations according to which the CGoal is formulated. Communicative obligations are shared by the agents belonging to the same communicative environment, and they appear to be rather persistent principles: even in the case of strong evidence that the partner does not share the agent's communication model, the agent may choose to change her application and world models rather than doubt the partner's communicative cooperation. CGoals contain several consistent intentions. The set can be augmented with intentions which deal with explanation or compensation, or some of the main attitudes can be dropped if not necessary in the context,
see [12]. Besides enabling the agent to produce different surface realisations from the same CGoal, the ltering of CGoal via communicative principles has a further advantage in that it allows contributions to be multifunctional, i.e. convey several several communicative goals simultaneously. Communicative contributions carry the speaker's beliefs and intentions, and when uttered in a speci c utterance event, not only the conventional meaning of the contribution is conveyed but the communicative context is changed as well. Contextual changes, however, can only be estimated as expectations of the partner's reaction, so the completion of a communicative act is usually achieved via a sequence of utterances where each contribution is a reaction to the previous contribution and speci es the context to the point where the agents agree that their CGoals have been successfully achieved (claims 7,8). Hence, a communicative goal is a complex kind of intention whose different e ects on the context are derived from general principles of rational agenthood and cooperative interaction. A similar view is also advocated by [7].
5 Conclusion
[5] C. Cherniak. Minimal Rationality. The MIT Press. Cambridge, Massachusetts, 1986. [6] H. H. Clark and D. Wilkes-Gibbs. Referring as a collaborative process. In P. R. Cohen, J. Morgan, and M. E. Pollack, editors, Intentions In Communication, pages 463{493. The MIT Press. Cambridge, Massachusetts, 1990. [7] P. R. Cohen and H. J. Levesque. Rational interaction as the basis for communication. In P. R. Cohen, J. Morgan, and M. E. Pollack, editors, Intentions in Communication, pages 221{255. The MIT Press. Cambridge, Massachusetts, 1990. [8] J. R. Galliers. A theoretical framework for computer models of cooperative dialogue, acknowledging multi-agent con ict. Technical Report 172, University of Cambridge, Computer Laboratory, 1989. [9] B. J. Grosz and C. L. Sidner. Attention, intentions, and the structure of discourse. Computational Linguistics, 12(3):175{204, 1986.
Meta-Dialogue Behaviors: Improving the Eciency of Human-Machine Dialogue. A Computational Model of Variable Initiative and Negotiation in Collaborative ProblemSolving. PhD thesis, Duke University, 1994. K. Jokinen. Response Planning in InformationSeeking Dialogues. PhD thesis, University of
[10] C. I. Guinn.
This paper has argued that advances in theoretical research are important even when building practical service systems. Although more research is needed to build large-scale dialogue systems that exhibit aspects of rationality in resource bounded agents, con ict reso- [11] lution etc., it is obvious that the new insights are going Manchester Institute of Science and Technology, to be an integral part of ambitious dialogue systems in 1994. future. Constructive Dialogue Model presents a new ap- [12] K. Jokinen. Communicative principles and utproach to dialogue management. The central claim is terance planning. In A. I. C. Monaghan, editor, that dialogues are managed locally by reacting to the CSNLP 1995. 4th International Conference on the changing dialogue context. Communicative principles Cognitive Science of Natural Language Processencode the agents' cooperativeness and rationality, and ing, Dublin City University, 1995. provide the basis for reasoning about dialogue contin[13] K. Jokinen. Rational agency. In M. Fehling, eduations appropriate to the changing context. itor, Rational Agency: Concepts, Theories, Models, and Applications, pages 89{93. Proceedings of References The AAAI Fall Symposium, MIT, Boston, 1995. [1] J. Allwood. Linguistic Communication as Action and Cooperation. Department of Linguistics, Uni- [14] J. D. Moore. Participating in Explanatory Diaversity of Goteborg, 1976. Gothenburg Monologues. Interpreting and Responding to Questions graphs in Linguistics 2. in Context. The MIT Press, Cambridge, Massachusetts, 1995. [2] D. Appelt. Planning Natural Language Utterances. Cambridge University Press, Cambridge, [15] J. Nivre. Corpus collection and analysis. Technical Report D2.1, PLUS deliverable., (Ed.)(1992. 1985. Studies in Natural Language Processing. [3] M. Bratman, D. Israel, and M. E. Pollack. Plans [16] W. Wahlster, H. Marburger, A. Jameson, and S. Buseman. Over-answering yes-no questions: and resource-bounded practical reasoning. ComExtended responses in a NL interface to a vision putational Intelligence, 4:349{355, 1988. system. In Proceedings of the 8th IJCAI Conference, pages 643{646, 1983. [4] A. Cawsey. Explanation and Interaction. The Computer Generation of Explanatory Dialogues. The MIT Press, Cambridge, Massachusetts, 1993.
